Relevant Information

Luciano Floridi—founder of to academic domains, philosophy of information and information ethics—spoke Friday at Indiana University’s School of Informatics about “A Subjectivist Interpretation of Relevant Information.” This is a work-in-progress trying to understand the philosophical model for relevance.

Philosophy of Information (PI) uses conceptual models to describe the dynamics and utility of information, and it also examines computational methods as a means to address philosophical problems (see Floridi’s defining 2002 paper). Since philosophy is the root of all science, this kind of relationship is not surprising. There are varied interpretation of what information is, but Floridi uses three dimensions to describe it. Information can be though of as reality (i.e. fingerprints), for reality (i.e. recipe), or about reality (descriptives). It is this third factual dimension on which his discussion of relevance is based.

Models of Relevant Information
Floridi’s current work is trying to model when information becomes relevant. Current theories are split into two types: system-based (“S-theories”) and agent-oriented (“A-theories”). They each approach information from different perspectives, the former dealing with measures like topicality, aboutness and covariance while the latter speaks of cognitive pertinence, perceived utility and beneficiality. S-theories are causal, and A-theories are epistemological. Neither, however, offers a definition of relevance. Instead, the underlying assumption is that there is such thing as relevant information, blackboxing the dynamics that might explain what it is. Thus, in the philosopher’s eyes, there is a need for a theory that does attempt to explain relevance.

Floridi has critiqued semantic information in the past (“Is Semantic Information Meaningful” PDF), and this relevant information model is an outgrowth of that. The basic case is an agent asking a question that receives an answer containing relevant information. The three basic components—agent, question, information—are contained within a specific domain, acknowledging the fact that good information in one context is not necessarily relevant in another. It is also agent-oriented in that the agent’s interest is embedded in the question she asks. However, there are problems with the basic model. There is no accounting for misinformation. Informativeness is indistinguishable from pertinence, so an answer could be technically relevant (“The train leaves from New York.”) but not answer the question (“When does the train leave?”). Also, there is no degree given to show how relevant something is.

Floridi attempts to address this last detriment by incorporating a sense of probability that the question and answer are adequate. Adequacy is measured in terms of accuracy and precision. An answer of high accuracy and low precision would be very specific to the question but not be very focused. An answer of high precision and low accuracy would be very concentrated on a specific topic, but that topic would be a better response to a different question. Adequacy of a question or an answer would have both high accuracy and high precision. Unfortunately, in the real world, an agent has no idea which questions are the best ones to ask, given that the world is opaque and much of its information is not available to the individual.

This gets into an old philosophical debate between Socrates and a student, Meno, who argued that inquiry is impossible. Either you know the answer to the question, and therefore don’t need to ask it, or you don’t know the question to ask, and can never get to the answer. Questions also aren’t always about getting new information of relevance, however, but rather confirming or reinforcing relevant information already in the agent’s possession. Inquiry becomes possible by asking questions you do know to bring equivalent new information that can inform future questions. Eventually, you will realize the question you need and ask it.

This problem is addressed by adding the concept of new information into the model. Given the presence of new information, an agent is likely to ask the question that will bring relevant answers. Or, conversely, the new information could inform the agent not to ask a misleading question.

Future Work (the relevance of Relevance)
What is missing right now are the behaviors that result from relevant information, and what happens if the new information is misleading. Behaviors may be a way of identifying misinformation, upon reflection, and therefore better assessing its relevance. In a complex world, the behaviors of an agent are directly tied to future interactions and thus changes the emergent properties of the system. That, in turn, can change the domain’s context, altering the relevance of the information. Bad information can lead to distractive questioning. Continuous bad information could lead to serendipity, where the agent arrives at the relevant answers through a series of irrelevant questions. Floridi acknowledges the youth of his model of relevant information. Issues like these may be incorporated as the model evolves.

While intensely abstract, this foundational thinking is very helpful in getting at the dynamics of information flow. My real-world problem deals with new members of a forum encountering a barrier in the number and organization of previous contributions. Summaries, which are significantly easier to digest than reading every thread in order, would help a new member understand the gestalt of the community and thus lower a barrier to entry. But what goes into that summary? What is the most relevant information for that person? If a solution to summarization were automated, the algorithm would have to incorporate some sense of relevancy to be able to present the most effective summary. If the summarization is collective, as in a wiki page maintained by discussants in a thread, the Floridi model might be ported to a complex-adaptive system simulation to help predict the dynamics of the community and optimize relevance.